source: http://github.com/muschellij2/CMStat_2018

Disclaimer (Acknowledgements?)

UK Biobank Data

  • Overall 500,000 participants (UK), ≈ 100,000 included in the sub-study

In this presentation:

  • Explore the data
  • Assess “bias” in different devices (see if “autocalibration” is working)
  • Also discuss inclusion criteria “bias”
  • Get similar findings (Doherty et al. 2017) analysis

Where do 100K come from?

Demographics: Lots of Non-Response

Completed: good data Completed: bad data No response Not asked
n 96701 7005 132800 266110
Age at Initial Visit (mean (sd)) 56.6 (7.8) 55.2 (7.9) 56.4 (8.0) 57.5 (8.1)
Male (% Male) 42255 (43.7) 3156 (45.1) 62601 (47.1) 121151 (45.5)
Ethnicity (% Non-White) 2983 (3.1) 335 (4.8) 7617 (5.8) 16102 (6.1)

Many people DIED before being able to be asked

Assessment to Accelerometry can be a WHILE

Responders are Healthier (Self-Reported)

Completed: good data Completed: bad data No response Not asked
Overall health (%)
Excellent 20987 (21.8) 1464 (21.0) 21583 (16.3) 37849 (14.4)
Good 57849 (60.0) 4057 (58.1) 78968 (59.7) 148196 (56.2)
Fair 15149 (15.7) 1261 (18.0) 26669 (20.2) 62313 (23.6)
Poor 2482 (2.6) 205 (2.9) 4969 (3.8) 15124 (5.7)

Lesson #1: The devil is
in the inclusion criteria
(or can be)

Data Gathered

  • Tri-axial Axtivity 100Hz over 7 days
  • Started at 10AM and ended at 10PM (spoiler: will be important)
  • Data measured in milli-g (1\(g\) = 9.80665 \(\frac{m}{s^2}\))
    • not counts or steps as other devices

Accelerometry Data Available

  • Data at varyling levels
    • Axtivity CWA format (Highest resolution, 100Hz) (200Mb per user)
      • very large for 100K subjects (20Tb)
  • 5 second level data
    • UKBB imputation/processing done
    • averaged into 1440 minute-level data
  • Overall statistics (mean/median): overall, daily, hourly, day of week
    • removed “non-wear” periods

Accelerometry Data Available

  • Data at varyling levels
    • Axtivity CWA format (Highest resolution, 100Hz) (200Mb per user)
      • very large for 100K subjects (20Tb)
  • 5 second level data
    • UKBB imputation/processing done
    • averaged into 1440 minute-level data
  • Overall statistics (mean/median): overall, daily, hourly, day of week
    • removed “non-wear” periods

UKBB Processing: Auto-calibration

(Hees et al. 2014)

  1. Use 10\(s\) window all axes SD \(< 13.0\) m\(g\).
  2. Fit a unit gravity sphere using OLS.
  3. If 3 axes had values outside a \(\pm 300\)m\(g\) range - use calibration coefficient
  4. If not, use next person’s calibration coefficient from the same device


Auto-calibration seems to “work”

  • Within-person average, within-device average (one point per device)
  • Plotted against # of wears per device(\(σ\) should decrease with \(\sqrt{n}\))

Doesn’t work for all cases

Doesn’t work for all cases

Doesn’t work for all cases

Lesson #2: If magnitude is important, need calibration
(“batch effect” correction),
but may not be perfect

UKBB Processing: Doherty et al. (2017)

  • Recording errors and ‘interrupts’ flagged (plug in accelerometer to computer)
  • \(\pm8g\) flagged
  • Resampled to 100 Hz (interrupts > 5 seconds set to missing)
  • Euclidean norm, fourth order Butterworth low pass filter (f = 20Hz).
  • Subtract \(1g\), negative values set to \(0\)

They have software (Python) on GitHub

One result from (Doherty et al. 2017) analysis

How do people typically move?

  • Average over minute - regardless of mulitple visits per person

How do people typically move?

  • Average over minute with days > 95% non-missing data

(Maybe) Lesson #3: Keep only “full” days

Not so fast, let’s look day by day

  • Average over minute for days separately, one row per subject

Removing Day 7

  • Average over minute for days 0 - 6

Lesson #3: Explore days before averaging across in individual?

Well use the MEDIAN then!

Maybe it’s a “few bad apples”

Heatmap of 2000 randomly sampled people

Age at Assessment Plot

One result from (Doherty et al. 2017) analysis

  • Look at age range

Age at Acceleration Plot

Age at Acceleration Plot - Remove 38-44 y/0

One result from (Doherty et al. 2017) analysis

Takehome Messages

  1. Start off smaller than 100K people
  2. Inspect the raw(ish) data
  3. Processing highly affects results
  • Autocalibration seems to work well on gross features (with 100K people)
    • Artifacts still seem present in the data
  1. Inclusion criteria matters (esp. for inference)
  2. Back to the 100Hz data we go!

Next data installment

“We invited some participants to wear an activity monitor for a week, four times a year. … finished in early 2019.”

References (and Thanks)

We’re ready for analysis right?

  • Nope! Need to estimate “wear time”

Wear time

“We removed non-wear time, defined as consecutive stationary episodes lasting for at least 60 minutes where all three axes had a standard deviation of less than 13.0 m\(g\)”

Doherty et al. (2017)

Why 13m\(g\)?

“Here, 13 m\(g\) was selected just above the empirically derived baseline (noise) standard deviation of 10 m\(g\) to retain only nonmovement periods.” Hees et al. (2014)

Doherty, Aiden, Dan Jackson, Nils Hammerla, Thomas Plötz, Patrick Olivier, Malcolm H Granat, Tom White, et al. 2017. “Large Scale Population Assessment of Physical Activity Using Wrist Worn Accelerometers: The Uk Biobank Study.” PloS One 12 (2): e0169649.

Hees, Vincent T van, Zhou Fang, Joss Langford, Felix Assah, Anwar Mohammad, Inacio CM da Silva, Michael I Trenell, Tom White, Nicholas J Wareham, and Søren Brage. 2014. “Autocalibration of Accelerometer Data for Free-Living Physical Activity Assessment Using Local Gravity and Temperature: An Evaluation on Four Continents.” Journal of Applied Physiology 117 (7): 738–44.